A thousand pages on your desk

Say you have eight agents running. Each one has been working all day, and each one has filled its 500k-token context window. That's roughly a 1,000-page book.

Eight books. Every day. Are you going to read them?

Of course you won't. So they pile up on your desk, and you feel dread.

We can accomplish much more with AI than we ever could without. But the way to run many agents is not to just do way more of whatever you were doing with a few agents.

It's a completely different job.

Startups have levels

I like saying startups have levels, where the level is the power of 2 that gets you to the headcount.

At level 1, it's you and a co-founder. You're on every decision, you understand every detail, you steer the product to exactly what you believe customers want. That's how you win at that size.

Every level up is a new ballgame. Around level 3 you start needing some loose management. At 4, real managers and departments. At 5, HR and operations. At 6, forecasts stop being just numbers on a spreadsheet.

By level 7, 128 people, you look back fondly at the days you reviewed every pull request and answered every support email yourself. Now you hear about launches at the all-hands.

The CEO of a 128-person company isn't doing 128 times the work of a solo founder. It's a different job.

Agents are the same. With one, you ask, it replies, you ask, it replies, and you're more productive than without it. With 5, you've already lost track of the details. With 50, you never had them.

So you need different signals. Not the 1,000-page books, but what executives get: briefs, pitches, presentations.

And you need to set up the group to succeed without you being involved in everything.

Now you have cloned yourself

Michael Hyatt's Free to Focus is from 2019, way before agents, but it has a great framework for this: five levels of delegation.

At level 1, the person does exactly what you asked, no more and no less. At level 2, they research and report back, and you decide. At level 3, they research and recommend, and you approve. At level 4, they decide, act, and tell you what they did. And then there's level 5:

At Level 5, you are effectively handing the entire project or task over to someone else and exiting the decision altogether. You'd say something like this: "Here's what I need you to do. Make whatever decision you think is best. There's no need to report back or tell me what you did." Now you have cloned yourself.

Hyatt, Michael. Free to Focus: A Total Productivity System to Achieve More by Doing Less (p. 153). Kindle Edition.

He wrote this about delegating to people. But it's never been more true than now, with agents.

Ask, reply, ask, reply is level 1. Most of us are trying to run a 128-person company with level 1 delegation.

Then it doesn't work, we burn out, and we blame the AI for being dumb.

AI is NOT a tool

Part of the problem is how we talk about it. We say "AI is a tool," that it's something you wield to get what you want, like a hammer or a spreadsheet.

Nope. You don't delegate to a hammer. You don't give a spreadsheet level 4 authority.

Working with AI isn't exactly like working with people, but it's a lot more like that than like working with tools. So stop treating it like one.

If you wouldn't work with someone by reading their mind, why are you reading your agent's context window? If you wouldn't follow 5 people around and make them stop and wait every time you leave the room, why do it with your agents?

Your job isn't to use the tool to do the work. Like a CEO, your job is to get the right AI for the job, set it up to succeed, and then get out of its way so the agent can do the work.

This place basically runs itself

There's a funny scene in The Morning Show, the Apple TV show. Chip is telling his girlfriend he'll take a day off to spend with her. Given he's a known workaholic, and TV news isn't exactly the poster child for work-life balance, she asks: "That'd be great, but are you sure you can free yourself up?"

"Yeah, this place basically runs itself."

Meanwhile, an employee comes over: "Hey boss, all the phone lines are down. None of the phones work."

"Hrm... okay. Have you tried calling the phone company?"

The employee, puzzled: "No. The phone lines are down."

"Hrm, can you use your mobile phone?"

"Ah! Yes." And he leaves.

Don't be like Chip.

You don't build a dark factory by going to your real factory and turning off the lights.

A team that runs smoothly while you're out doesn't get that way by accident. Someone built it to that level of autonomy.

That also applies to agents. You can't just prompt any agent to do any work, walk away, hope for the best, and get angry when it doesn't work.

Building a setup where many agents do great work is like building a great company with many employees.

It costs time.

And it costs money.

And there's no way around it.

Fixing the verification bottleneck

There's a huge asymmetry between the typical human effort needed to create work with AI and the effort to verify it.

If you one-shot prompt Fable 5.2 to build an FPS game, it will. But how do you verify the game is "right," by whatever your definition of right is? Can you one-shot that? Is "make no mistakes" going to apply the exact verification you'd expect?

No.

So you're the bottleneck. And Goldratt's Theory of Constraints says that producing more before the bottleneck doesn't get more done. It just piles up work waiting in front of it.

So the human work can't be verification. It doesn't scale. What you do instead is the work to reduce the asymmetry between how much human effort it takes to create and how much human effort it takes to verify.

Otherwise you're going to burn out while drowning in verification.

But knowing the theory doesn't get you off the hook. You can't go to your customers and say "stop asking me for stuff, you're violating Goldratt!!" The demand keeps coming.

You have to fix your production line while you keep it running.

Burnout

The problem with burnout is a lack of control over inputs.

ICs feel it less, because their manager controls their input (even if it doesn't seem like it sometimes). Even on a death march, someone else decides what lands on their plate. Leaders don't get that, especially at startups. Your manager might ask "hey, are you free to work on this thing?" Your customer won't. And neither will your agents.

That's the dread from the start. It isn't just the amount of work. Eight agents will produce more than you can ever read, and nothing in the system says "that's enough for today." Except you. And even if you could read all those 1,000-page books, they'd just produce more. It's unwinnable.

And it gets worse, because it's not just more work. It's harder work. Leaders, for better or (probably) for worse, only get the worst work. Whatever AI can do easily, it will. What reaches you is the hard work, the hard decisions, the stuff that has to go through you for the whole thing to get done.

As your setup matures, AI takes that work too. Then what's left for you is even harder.

You could choose to stop pushing. Build a machine that solves problems of a certain size, like a lifestyle business for agentic workflows, and just watch it run. If all you wanted from AI was writing LinkedIn posts, that'd be pretty easy to set up today.

But if you're trying to have AI do your whole job, and you're competing with other people who are also using AI to do your whole job, you have no choice but to keep pushing to outcompete them.

So, how to work with 40+ agents?

I don't know.

But when I'm working with 40 agents, I sure as hell am not reading 40 context windows.

I'm reading signals. Are they working or idle? Is the work in progress or done? Are they talking to each other? Waiting? Do they seem stuck?

Are the tests passing? Do I trust the tests to ensure the software works? Did it break even though the tests passed? Why? How can I fix it for that whole class of problems in the future?

And when I'm not around, I have a flag that says "delegate to my proxy" on my control plane: my proxy is being trained and gets better and better at making decisions the way I would.

But not always. And then, how can I help it make better decisions in the future?

How can I equip it to succeed without me, instead of with me?

That's the job.

And it's hard.